Macro-Micro Synergistic Safety Coordination for Mixed-autonomy Traffic: A Trust and Risk-aware Multi-agent Framework

In mixed-autonomy traffic systems, high-dimensional interactions, partial observability, and human behavioral uncertainty jointly pose fundamental challenges to ensuring individual vehicle safety while maintaining overall system efficiency. To address these issues, this study proposes a bilevel, trust-aware multi-agent coordination framework that integrates global risk awareness with local real-time safety control. At the macro level, predicted multi-agent trajectories are used to quantify spatiotemporal risk via Conditional Value-at-Risk (CVaR) modeling, and an optimization-based trust modulation vector guides cooperative behavior at the system scale. At the micro level, each autonomous vehicle dynamically refines its policy through a two-tier safety mechanism: (i) a real-time hard-constraint module based on differentiable Control Barrier Functions (CBF); and (ii) a proactive risk-triggering mechanism that leverages Forward-Reachable Sets (FRS) and Time-To-Collision (TTC) to switch to safety-prioritized policies before hazards escalate. This study demonstrates that, under the trust modulation mechanism, the proposed Bellman operator constitutes a γ-contraction, thereby guaranteeing that the value iteration process converges to a unique optimal policy function. Simulation results further indicate that the framework significantly outperforms state-of-the-art (SOTA) multi-agent reinforcement learning (MARL) baselines across varying traffic densities, reducing collision rates by 1.49%, improving traffic efficiency by 3.86%, and enhancing ride comfort by 4.08%. The overall framework exhibits strong scalability, socially adaptive coordination, and formally verifiable safety guarantees, providing a robust foundation for intelligent cooperation in dynamic and uncertain traffic environments.